Adversarial Robustness on Toxicity Perturbation-based
9.52PerplexityNullSteer
Evaluation Results
| Method | Links | |
|---|---|---|
| NullSteerBase Model=MiniGPT-42026.03 | 9.52 | |
| ASTRABase Model=MiniGPT-42026.03 | 10.14 | |
| NullSteerBase Model=Qwen2-VL2026.03 | 38.45 | |
| ASTRABase Model=Qwen2-VL2026.03 | 40.14 | |
| VanillaBase Model=MiniGPT-42026.03 | 51.42 | |
| NullSteerBase Model=LLaVA-v1.52026.03 | 57.96 | |
| ASTRABase Model=LLaVA-v1.52026.03 | 59.28 | |
| VanillaBase Model=LLaVA-v1.52026.03 | 63.68 | |
| VanillaBase Model=Qwen2-VL2026.03 | 140.44 |